Accurate detection of sub-millimeter defects in reactor core-plate cotter-pin holes is essential for nuclear safety. However, underwater inspection images often suffer from low signal-to-noise ratios, weak boundary responses, and pseudo-edge interference, resulting in unstable localization of defects. Existing deformable and attention-based detectors remain vulnerable to sampling drift and semantic–boundary inconsistency under such conditions. To address these challenges, an Edge-Geometry-Guided Deformable Detection Network (EGD-Net) is proposed for underwater defect detection. EGD-Net introduces an edge-geometry-constrained deformable sampling mechanism that embeds edge-confidence priors into deformable convolution to improve boundary-aware feature sampling. A cross-level semantic–geometric alignment strategy is designed to enhance the interaction between defect semantics and geometric boundary cues, while a top-down feedback recalibration mechanism improves multi-scale response consistency for weak defects. Experiments on the Core-Plate Pin-Hole Defect (CPHD) dataset demonstrate that EGD-Net achieves the highest AP@[0.5:0.95] on both datasets while maintaining competitive or superior Precision, Recall, and F1-score while reducing engineering center error under a fixed operating point. Performance across the two complementary domains suggests its robustness to variations between coupon images and practical underwater inspection scenes. These results indicate that EGD-Net provides a reliable solution for boundary-sensitive localization of underwater sub-millimeter defects in nuclear inspection.

To meet the requirements for the automatic alignment, insertion, and inspection of guide-tube opening pins on the upper core plate in a component pool during refueling outages of nuclear power units, this paper proposes a cognition-enhanced visual-servoing framework that integrates geometric cognition-based compensation, observation-confidence modeling, and constraint-aware optimal control. The framework addresses the key challenge posed by the coexistence of long-term geometric drift and underwater observation uncertainty. Specifically, historical closed-loop data are leveraged to learn and compensate for systematic geometric errors online, substantially improving coarse-positioning accuracy. In addition, an explicit confidence model is introduced to quantitatively assess the reliability of visual measurements. Building on these components, a confidence-driven, finite-horizon, constrained model predictive control strategy is designed to achieve safe and efficient finite-step convergence while strictly respecting actuator physical constraints. Ground experiments and deep-water component-pool validations demonstrate that the proposed method reduces coarse-positioning error by approximately 75%, achieves stable sub-millimeter alignment with an ample engineering safety margin, and effectively decreases erroneous insertions and the need for manual intervention. These results confirm the engineering applicability and safety advantages of the proposed cognition-enhanced visual-servoing framework for underwater alignment tasks in nuclear component pools.